123         ap += 4;  in bn_mul_add_words()
124         rp += 4;  in bn_mul_add_words()
125         num -= 4;  in bn_mul_add_words()
153         ap += 4;  in bn_mul_words()
154         rp += 4;  in bn_mul_words()
155         num -= 4;  in bn_mul_words()
177         sqr(r[4], r[5], a[2]);  in bn_sqr_words()
179         a += 4;  in bn_sqr_words()
181         n -= 4;  in bn_sqr_words()
190         sqr(r[4], r[5], a[2]);  in bn_sqr_words()
198  asm("divq      %4":"=a"(ret), "=d"(waste)  in bn_div_words()
216                   ".p2align 4                     \n"  in bn_add_words()
217                   "1:     movq    (%4,%2,8),%0    \n"  in bn_add_words()
243                   ".p2align 4                     \n"  in bn_sub_words()
244                   "1:     movq    (%4,%2,8),%0    \n"  in bn_sub_words()
301         a += 4;  in bn_sub_words()
302         b += 4;  in bn_sub_words()
303         r += 4;  in bn_sub_words()
309 /* mul_add_c(a,b,c0,c1,c2)  -- c+=a*b for three word number c=(c2,c1,c0) */
310 /* mul_add_c2(a,b,c0,c1,c2) -- c+=2*a*b for three word number c=(c2,c1,c0) */
311 /* sqr_add_c(a,i,c0,c1,c2)  -- c+=a[i]^2 for three word number c=(c2,c1,c0) */
313  * sqr_add_c2(a,i,c0,c1,c2) -- c+=2*a[i]*a[j] for three word number
314  * c=(c2,c1,c0)
323 #  define mul_add_c(a,b,c0,c1,c2)       do {    \
327         c0 += lo; hi += (c0<lo)?1:0;            \
331 #  define mul_add_c2(a,b,c0,c1,c2)      do {    \
335         c0 += lo; tt = hi+((c0<lo)?1:0);        \
337         c0 += lo; hi += (c0<lo)?1:0;            \
341 #  define sqr_add_c(a,i,c0,c1,c2)       do {    \
345         c0 += lo; hi += (c0<lo)?1:0;            \
349 #  define mul_add_c(a,b,c0,c1,c2) do {  \  argument
355         asm ("addq %3,%0; adcq %4,%1; adcq %5,%2"       \
356                 : "+r"(c0),"+r"(c1),"+r"(c2)            \
361 #  define sqr_add_c(a,i,c0,c1,c2) do {  \  argument
367         asm ("addq %3,%0; adcq %4,%1; adcq %5,%2"       \
368                 : "+r"(c0),"+r"(c1),"+r"(c2)            \
373 #  define mul_add_c2(a,b,c0,c1,c2) do { \  argument
379         asm ("addq %3,%0; adcq %4,%1; adcq %5,%2"       \
380                 : "+r"(c0),"+r"(c1),"+r"(c2)            \
383         asm ("addq %3,%0; adcq %4,%1; adcq %5,%2"       \
384                 : "+r"(c0),"+r"(c1),"+r"(c2)            \
390 # define sqr_add_c2(a,i,j,c0,c1,c2)      \  argument
391         mul_add_c2((a)[i],(a)[j],c0,c1,c2)
418     mul_add_c(a[4], b[0], c2, c3, c1);  in bn_mul_comba8()
422     mul_add_c(a[0], b[4], c2, c3, c1);  in bn_mul_comba8()
423     r[4] = c2;  in bn_mul_comba8()
426     mul_add_c(a[1], b[4], c3, c1, c2);  in bn_mul_comba8()
429     mul_add_c(a[4], b[1], c3, c1, c2);  in bn_mul_comba8()
435     mul_add_c(a[4], b[2], c1, c2, c3);  in bn_mul_comba8()
437     mul_add_c(a[2], b[4], c1, c2, c3);  in bn_mul_comba8()
445     mul_add_c(a[3], b[4], c2, c3, c1);  in bn_mul_comba8()
446     mul_add_c(a[4], b[3], c2, c3, c1);  in bn_mul_comba8()
455     mul_add_c(a[4], b[4], c3, c1, c2);  in bn_mul_comba8()
463     mul_add_c(a[4], b[5], c1, c2, c3);  in bn_mul_comba8()
464     mul_add_c(a[5], b[4], c1, c2, c3);  in bn_mul_comba8()
470     mul_add_c(a[6], b[4], c2, c3, c1);  in bn_mul_comba8()
472     mul_add_c(a[4], b[6], c2, c3, c1);  in bn_mul_comba8()
476     mul_add_c(a[4], b[7], c3, c1, c2);  in bn_mul_comba8()
479     mul_add_c(a[7], b[4], c3, c1, c2);  in bn_mul_comba8()
524     r[4] = c2;  in bn_mul_comba4()
558     sqr_add_c2(a, 4, 0, c2, c3, c1);  in bn_sqr_comba8()
559     r[4] = c2;  in bn_sqr_comba8()
562     sqr_add_c2(a, 4, 1, c3, c1, c2);  in bn_sqr_comba8()
567     sqr_add_c2(a, 4, 2, c1, c2, c3);  in bn_sqr_comba8()
575     sqr_add_c2(a, 4, 3, c2, c3, c1);  in bn_sqr_comba8()
578     sqr_add_c(a, 4, c3, c1, c2);  in bn_sqr_comba8()
586     sqr_add_c2(a, 5, 4, c1, c2, c3);  in bn_sqr_comba8()
590     sqr_add_c2(a, 6, 4, c2, c3, c1);  in bn_sqr_comba8()
594     sqr_add_c2(a, 7, 4, c3, c1, c2);  in bn_sqr_comba8()
633     r[4] = c2;  in bn_sqr_comba4()